Agents Reviewing Agent-Generated Code Is Either Brilliant or a House of Cards

Anthropic launched Claude Code Review, a multi-agent system that dispatches parallel AI reviewers to check pull requests. Each agent examines a different issue type — security, logic errors, style violations — then findings get verified and ranked by severity before surfacing to the developer.

The internal numbers from Anthropic’s own usage are striking. Before Code Review, only 16% of PRs received substantive review comments. After: 54%. Engineers marked less than 1% of findings as incorrect. Large PRs with 1,000+ changed lines received findings 84% of the time.

The price tag: $15-$25 per PR in token usage.

The self-referential loop here deserves attention. AI generates more code than humans can review, so we build AI to review the AI-generated code. This is either a virtuous cycle that raises code quality across the board, or a system where the same blind spots compound undetected.

For now, the evidence points toward virtuous cycle. A multi-agent approach where specialized reviewers each focus on different failure modes is genuinely harder to fool than a single-pass review. Verification of findings before surfacing them reduces noise. And $15-$25 per PR is cheap compared to the cost of a production bug that slipped through a 2,000-line PR nobody had time to read properly.

But here’s the integration question that matters more than the capability one: does your team actually act on review findings?

Most engineering organizations already struggle with human code review comments. Reviews sit unaddressed. Comments get acknowledged but not fixed. “Will address in a follow-up” becomes permanent technical debt.

An AI that catches issues in 84% of large PRs is useless if:

• Nobody is accountable for resolving findings
• The review output gets treated as optional
• Teams lack the process to triage AI-generated feedback alongside human feedback

The tool works. The question is whether your organization has the review culture to make it matter. A $25-per-PR investment that nobody reads is just a more expensive way to ignore the same problems.

What does your team’s PR review completion rate look like today — before you add another source of findings to the queue?